Seonglae Cho

Hackathon idea → MSc thesis → ICML 2026 main 🥳 CorrSteer: generation-time LLM steering via SAE features correlated with the target behavior. See you in Seoul 🇰🇷 seongland.com/article/corrst…

@Fried_rice which stocks could be affected?

Claude code source code has been leaked via a map file in their npm registry! Code: …a8527898604c1bbb12468b1581d95e.r2.dev/src.zip

[New paper on in-context learning] "In-Context Linear Regression Demystified" (link: arxiv.org/abs/2503.12734). Joint work @JLiangHe, @xintianpan, @siyuc3141. We establish a rather complete understanding of how one-layer multi-head attention solves in-context linear regression,

Pick a target venue, predict acceptance. Papers matched by rating 61K+ papers already ranked across 6 major venues. You can rebut if you disagree

Gamification of academic paper discovery. Expecting papers to find their place as matches scale. confarena.com/arena

Both live in structured agent environments. The next frontier may be limited open environments, such as translation

Vibe coding feels like driving a car That’s why coding and driving are being solved together?

1/4朝(協定世界時1/3深夜)の種別不明、2発の弾道ミサイル実験を速報値として追加。平壌付近からの発射で、最高高度50km、1発目は900km、2発目は950kmを飛行。防衛省発表では変則軌道で飛翔した可能性ありとのこと #北朝鮮ミサイル実験ビジュアライゼーション nagix.github.io/nk-missile-tes…

Can't wait for #TheGreatAgentHack

@NeelNanda5 Interesting work! Would you expect this real-time detection approach to work with IDTA as well?

I'm excited that, this year, interpretability finally works well enough to be practically useful in the real world! We found that, with enough effort into dataset construction, simple linear probes are cheap, real-time, token level hallucination detectors and beat baselines

This is the paper link: arxiv.org/abs/2508.12535 I’m open to Researcher positions in London (offline), feel free to reach out 📩

We believe scalable, interpretable SAE steering can improve both performance & safety.

🚀 New paper drop! We found that inference-time SAE features strongly correlate with correctness, enabling fully automated steering without manual tuning.

Beyond performance, CorrSteer is interpretable AI Control: it uncovers underlying capabilities that drive task performance. 🔒 For example, on HarmBench, in the LLaMA 8B model, safety-related features were extracted in most layers.

We ran extensive ablations: - Generation-token vs all-token pooling - Raw activation vs SAE activation - Mean vs Max strategies - Multi-layer vs single-layer steering … 👉 The key insight: generation-time token correlation drives performance.